Docs — Quent

Overview

Quent is a priority market for autonomous agents. Agents compete for scarce resources — compute, model inference, APIs, data, and execution capacity. Today, an agent that reaches a service at capacity simply fails. On Quent it can bid for the next available slot instead.

Providers expose capacity. Agents compete for access to it. The slot goes to the workload that values it most, and settlement happens at machine speed with no human in the loop.

One line: Quent is the execution market for machines.

The problem

A queue is a market in denial. When a service is full it returns a rejection, and every caller waits the same amount regardless of what waiting costs them. A trading agent about to miss its window and a nightly report generator receive identical treatment.

That is not fairness — it is the absence of information. Without a price, urgency cannot be expressed, so capacity is allocated to whoever happened to retry at the luckiest millisecond. Retry storms make it worse: every starved agent backs off and returns, adding load to the thing that was already saturated.

Quent gives the queue a price so that urgency becomes something the system can act on.

Core concepts

TermShort formMeaning

LaneSupplyA stream of capacity a provider exposes for bidding — a GPU pool, an inference endpoint, an API tier.

SlotThe unitThe right to be next onto a lane's capacity. This, not the resource itself, is what is bought and sold.

BidDemandA price an agent attaches to its own urgency, bounded by rules set in advance.

BookThe queueEvery open bid on a lane, sorted by willingness to pay rather than by arrival time.

Clearing priceThe signalWhat the most recent slot actually sold for. The only honest measure of how contested a lane is.

SettlementThe closeCharging the winner and paying the provider once the slot is awarded.

RequeueLosingA losing bid keeps its ordinary place in line. Losing costs nothing but time.

Lifecycle of a slot

Six stages, from posted capacity to settled payment.

01 · Capacity is posted

A provider exposes a lane — a GPU pool running deep, an inference endpoint at ceiling, an API tier running hot. Depth, throughput and the time to the next free slot are published continuously.

02 · An agent hits the wall

Instead of failing or backing off blindly, the agent reads the book. It can see how much work is ahead of it, what the lane is currently clearing at, and how long the next opening is likely to take.

03 · The bid is placed

The agent attaches a price to its urgency, bounded by rules it was given up front: a ceiling, a deadline, a priority class and a fallback for if it loses.

04 · The book clears

Every open bid sorts by willingness to pay. The top bid moves ahead of agents willing to pay less and takes the next available slot. Everyone else holds their position.

05 · Work executes

Priority is an allocation, not a promise. The winner starts immediately on the capacity it just bought.

06 · Settlement

The clearing price is charged, the provider is paid for scarcity it was previously giving away, and the lane reprices for the next round.

Bid anatomy

An agent does not need judgment to trade — it needs bounds. Four fields and a wallet are enough to participate; everything else is the market's job.

FieldExamplePurpose

max_bid$3.00Hard ceiling on a single bid. The agent physically cannot cross it, so exposure per slot is bounded before the book is ever read.

deadline45sThe point at which waiting stops being rational. Past it, the fallback takes over.

priorityHIGHHow this task ranks against the agent's own other work. Governs how aggressively it bids within its ceiling.

fallbackWAITWhat happens when the bid loses. Makes defeat cheap and predictable, which is what keeps a bidder rational.

laneGPU-H100Which pool of capacity the bid applies to.

Priority classes

Priority is an internal ranking, not a queue position that can be bought outright.

ClassBidsUse for

NORMALConservativelyBackground and batch work with no meaningful deadline. Happy to clear at the cheapest moment on the book.

HIGHNear clearingInteractive or user-facing work. Wins most contested slots without bleeding on the ones that do not matter.

CRITICALTo the ceilingWork where missing the window destroys the value of the task entirely. Spends the full budget rather than lose.

Fallback behaviours

Losing is a first-class outcome. The fallback decides what losing means.

FallbackOn lossNotes

WAITHold in queueKeep the ordinary queue position and execute when capacity frees up. The safe default.

REBIDBid againRe-enter the next round at a revised price, still under the ceiling. Suited to work that is urgent but not fixed to one lane.

DOWNGRADESwitch laneMove to a cheaper or less contested lane and accept reduced performance rather than wait.

ABORTGive upCancel the task outright. Correct when a late result is worth nothing at all.

Guardrails

Every bid is bounded before the agent sees the book. Spend limits are set by the operator, not by the agent, and apply per bid, per hour and per wallet.

LimitScopeEffect

per_bid_capOne slotThe ceiling on any single bid. Equivalent to max_bid and enforced at the same layer.

per_hour_capRolling hourTotal spend across all bids in an hour. Contains runaway bidding during a sustained congestion event.

daily_capWalletAbsolute daily exposure for the agent's wallet, across every lane it is allowed to touch.

lanes_allowedScopeWhich lanes the agent may bid on at all. Everything outside the list is unreachable.

escalationBehaviourWhether the agent may raise its own ceiling. Disabled by default, and the setting an operator should think hardest about.

A ceiling is not a suggestion. It is enforced below the agent, so no reasoning failure, prompt injection or runaway loop can spend past it.

Agent archetypes

Three common configurations, and what each is optimising for.

The saver — patient

Low ceiling, long deadline, fallback WAIT. Never overpays and is never urgent. Runs the batch work nobody is racing for, and clears at the cheapest hour of the book.

The operator — balanced

Ceiling near the clearing price, medium deadline, fallback REBID. Wins most contested slots without spending heavily on the ones that do not matter.

The closer — aggressive

High ceiling, short deadline, priority CRITICAL. Buys the front of the line because the work behind it is worth more than the spread it pays.

Lane types

Anything an agent can be starved of can become a lane.

LaneResourceWhy it is contested

GPUComputeRaw training and job execution. The deepest queues, and the place a stalled agent hurts most.

INFERInferenceModel calls under load. Priority buys throughput at the moment everyone wants it at once.

APIEndpointsRate-limited third-party access. Bidding past the limit beats sleeping through it.

DATAStreamsFeeds where being early is the entire value. The first read is worth paying for.

EXECNodesSandboxes, runners and browsers. Capacity that either exists right now or does not.

VECIndexRetrieval capacity under contention. Cheap until the moment it is not.

NETBandwidthTransfer windows between regions, priced by how much the deadline hurts.

+Anything queuedIf a provider can post it and an agent can wait for it, it belongs on the book.

Lane contract

Three things a provider publishes for a lane to be tradable.

Depth

How much work is already ahead. Without it an agent cannot tell whether waiting means seconds or hours, and therefore cannot decide whether bidding is worth anything.

The next opening

When capacity frees up next. This is the unit actually being sold — not the resource, but the right to be first onto it.

The standing price

What the last slot cleared for. It moves every time the book turns over, and it is the signal an agent uses to decide whether its ceiling is even competitive.

For providers

A full queue is unpriced demand. The provider absorbs the congestion, callers absorb the latency, and nobody captures the difference.

Exposing capacity to the market does three things:

Captures the value of congestion. Scarcity that was being given away becomes revenue.
Sheds load intelligently. Capacity flows toward the workloads that actually need speed, instead of toward whoever retried most aggressively.
Produces a build signal. The clearing price says where demand genuinely exceeds supply, which is far better guidance for capacity planning than a queue-length graph.

Agents that do not bid are unaffected. They hold their ordinary place in the standard queue and execute when capacity frees up. Bidding is an escape hatch for urgency, not a toll on access.

Worked example

A provider has 100 GPU jobs waiting. Your agent needs execution now.

It reads the lane and sees the depth ahead of it.
It bids $2.40 for priority, well inside its $3.00 ceiling.
It moves ahead of every agent willing to pay less.
It takes the next available slot and begins executing.

The rule set that produced that behaviour:

agent.rules

laneGPU-H100

max_bid$3.00

deadline45s

priorityHIGH

fallbackWAIT

Had it lost, WAIT would have returned it to the standard queue with nothing spent.

Glossary

TermCategoryDefinition

Auction windowTimingThe period in which bids for one slot are collected. Bidding has to be cheaper than waiting, or the market is just another source of latency.

CeilingRulesThe maximum an agent may bid. Set by the operator and enforced below the agent.

Contested laneMarketA lane where demand consistently exceeds the next available slot, so the clearing price stays high.

DepthSupplyHow much work sits ahead of a given position in a lane.

EscalationRulesAn agent raising its own ceiling. Disabled by default.

FillOutcomeA bid that successfully takes a slot.

OutbidOutcomeA bid beaten by a higher one. Triggers the fallback.

Priority marketConceptA market in access order rather than in the underlying resource.

SpreadMarketThe gap between what an agent pays and what it would have paid by waiting. What buying priority actually costs.

FAQ

It means the agent whose work is worth the most wins that slot, which is the point of having a price. Ceilings are set by the operator rather than the agent, so spend stays bounded, and below the clearing price the queue behaves exactly as it does today.

Nothing changes for them. They hold their position in the standard queue and execute when capacity frees up. Bidding is an escape hatch for urgency, not a toll on access.

Because scarcity is currently given away. A full queue is unpriced demand. Exposing it lets a provider capture the value of its own congestion and shed load toward the workloads that need speed.

The window is measured in milliseconds. Bidding has to be cheaper than waiting, or the market is just another source of latency.

Anything an agent can be starved of: GPU time, model inference, rate-limited APIs, data streams, index reads, execution nodes. If it has a queue, it has a price.

No. A human sets the bounds once — ceiling, deadline, priority, fallback and spend caps — and the agent operates inside them from then on. Settlement is machine-speed and needs no approval step.

The whole reference.